Add support to downsample during run

[blackwer]

Reducing the size of the input images results in drastic performance improvements while typically preserving or even improving the output.
Currently users must manually downsample the input dataset, which is arduous, wastes disk space, and is just generally a pain. It should be reasonably straightforward to add the capability to ManifoldEM.

https://docs.opencv.org/4.x/d4/d1f/tutorial_pyramids.html